mnist.py 9.47 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
#  Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
#  Licensed under the Apache License, Version 2.0 (the "License");
#  you may not use this file except in compliance with the License.
#  You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
#  Unless required by applicable law or agreed to in writing, software
#  distributed under the License is distributed on an "AS IS" BASIS,
#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#  See the License for the specific language governing permissions and
#  limitations under the License.
"""Convolutional Neural Network Estimator for MNIST, built with tf.layers."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import argparse
import os
22
import sys
23
24

import tensorflow as tf
25
import dataset
26
27


Asim Shankar's avatar
Asim Shankar committed
28
class Model(tf.keras.Model):
Asim Shankar's avatar
Asim Shankar committed
29
  """Model to recognize digits in the MNIST dataset.
Asim Shankar's avatar
Asim Shankar committed
30
31
32
33
34
35

  Network structure is equivalent to:
  https://github.com/tensorflow/tensorflow/blob/r1.5/tensorflow/examples/tutorials/mnist/mnist_deep.py
  and
  https://github.com/tensorflow/models/blob/master/tutorials/image/mnist/convolutional.py

Asim Shankar's avatar
Asim Shankar committed
36
  But written as a tf.keras.Model using the tf.layers API.
Asim Shankar's avatar
Asim Shankar committed
37
  """
Asim Shankar's avatar
Asim Shankar committed
38
39
40
41
42
43
44
45
46
47

  def __init__(self, data_format):
    """Creates a model for classifying a hand-written digit.

    Args:
      data_format: Either 'channels_first' or 'channels_last'.
        'channels_first' is typically faster on GPUs while 'channels_last' is
        typically faster on CPUs. See
        https://www.tensorflow.org/performance/performance_guide#data_formats
    """
Asim Shankar's avatar
Asim Shankar committed
48
    super(Model, self).__init__()
Asim Shankar's avatar
Asim Shankar committed
49
50
51
52
53
54
55
56
57
58
59
60
    if data_format == 'channels_first':
      self._input_shape = [-1, 1, 28, 28]
    else:
      assert data_format == 'channels_last'
      self._input_shape = [-1, 28, 28, 1]

    self.conv1 = tf.layers.Conv2D(
        32, 5, padding='same', data_format=data_format, activation=tf.nn.relu)
    self.conv2 = tf.layers.Conv2D(
        64, 5, padding='same', data_format=data_format, activation=tf.nn.relu)
    self.fc1 = tf.layers.Dense(1024, activation=tf.nn.relu)
    self.fc2 = tf.layers.Dense(10)
61
    self.dropout = tf.layers.Dropout(0.4)
Asim Shankar's avatar
Asim Shankar committed
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
    self.max_pool2d = tf.layers.MaxPooling2D(
        (2, 2), (2, 2), padding='same', data_format=data_format)

  def __call__(self, inputs, training):
    """Add operations to classify a batch of input images.

    Args:
      inputs: A Tensor representing a batch of input images.
      training: A boolean. Set to True to add operations required only when
        training the classifier.

    Returns:
      A logits Tensor with shape [<batch_size>, 10].
    """
    y = tf.reshape(inputs, self._input_shape)
    y = self.conv1(y)
    y = self.max_pool2d(y)
    y = self.conv2(y)
    y = self.max_pool2d(y)
    y = tf.layers.flatten(y)
    y = self.fc1(y)
    y = self.dropout(y, training=training)
    return self.fc2(y)


def model_fn(features, labels, mode, params):
  """The model_fn argument for creating an Estimator."""
  model = Model(params['data_format'])
90
91
92
93
  image = features
  if isinstance(image, dict):
    image = features['image']

Asim Shankar's avatar
Asim Shankar committed
94
  if mode == tf.estimator.ModeKeys.PREDICT:
95
96
97
98
99
100
101
102
103
104
105
    logits = model(image, training=False)
    predictions = {
        'classes': tf.argmax(logits, axis=1),
        'probabilities': tf.nn.softmax(logits),
    }
    return tf.estimator.EstimatorSpec(
        mode=tf.estimator.ModeKeys.PREDICT,
        predictions=predictions,
        export_outputs={
            'classify': tf.estimator.export.PredictOutput(predictions)
        })
Asim Shankar's avatar
Asim Shankar committed
106
  if mode == tf.estimator.ModeKeys.TRAIN:
107
    optimizer = tf.train.AdamOptimizer(learning_rate=1e-4)
108
109
110
111
112

    # If we are running multi-GPU, we need to wrap the optimizer.
    if params.get('multi_gpu'):
      optimizer = tf.contrib.estimator.TowerOptimizer(optimizer)

113
    logits = model(image, training=True)
114
    loss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits)
115
    accuracy = tf.metrics.accuracy(
116
        labels=labels, predictions=tf.argmax(logits, axis=1))
117
118
119
120
121
122
123
124
    # Name the accuracy tensor 'train_accuracy' to demonstrate the
    # LoggingTensorHook.
    tf.identity(accuracy[1], name='train_accuracy')
    tf.summary.scalar('train_accuracy', accuracy[1])
    return tf.estimator.EstimatorSpec(
        mode=tf.estimator.ModeKeys.TRAIN,
        loss=loss,
        train_op=optimizer.minimize(loss, tf.train.get_or_create_global_step()))
Asim Shankar's avatar
Asim Shankar committed
125
  if mode == tf.estimator.ModeKeys.EVAL:
126
    logits = model(image, training=False)
127
    loss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits)
128
129
130
131
132
133
    return tf.estimator.EstimatorSpec(
        mode=tf.estimator.ModeKeys.EVAL,
        loss=loss,
        eval_metric_ops={
            'accuracy':
                tf.metrics.accuracy(
Mark Daoust's avatar
Mark Daoust committed
134
                    labels=labels,
135
136
                    predictions=tf.argmax(logits, axis=1)),
        })
137
138


139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
def validate_batch_size_for_multi_gpu(batch_size):
  """For multi-gpu, batch-size must be a multiple of the number of
  available GPUs.

  Note that this should eventually be handled by replicate_model_fn
  directly. Multi-GPU support is currently experimental, however,
  so doing the work here until that feature is in place.
  """
  from tensorflow.python.client import device_lib

  local_device_protos = device_lib.list_local_devices()
  num_gpus = sum([1 for d in local_device_protos if d.device_type == 'GPU'])
  if not num_gpus:
    raise ValueError('Multi-GPU mode was specified, but no GPUs '
      'were found. To use CPU, run without --multi_gpu.')
154

155
156
157
158
159
160
161
162
163
  remainder = batch_size % num_gpus
  if remainder:
    err = ('When running with multiple GPUs, batch size '
      'must be a multiple of the number of available GPUs. '
      'Found {} GPUs with a batch size of {}; try --batch_size={} instead.'
      ).format(num_gpus, batch_size, batch_size - remainder)
    raise ValueError(err)


164
def main(unused_argv):
165
166
167
168
169
170
171
172
173
174
175
  model_function = model_fn

  if FLAGS.multi_gpu:
    validate_batch_size_for_multi_gpu(FLAGS.batch_size)

    # There are two steps required if using multi-GPU: (1) wrap the model_fn,
    # and (2) wrap the optimizer. The first happens here, and (2) happens
    # in the model_fn itself when the optimizer is defined.
    model_function = tf.contrib.estimator.replicate_model_fn(
        model_fn, loss_reduction=tf.losses.Reduction.MEAN)

Asim Shankar's avatar
Asim Shankar committed
176
177
178
179
  data_format = FLAGS.data_format
  if data_format is None:
    data_format = ('channels_first'
                   if tf.test.is_built_with_cuda() else 'channels_last')
180
  mnist_classifier = tf.estimator.Estimator(
181
      model_fn=model_function,
Asim Shankar's avatar
Asim Shankar committed
182
183
      model_dir=FLAGS.model_dir,
      params={
184
185
          'data_format': data_format,
          'multi_gpu': FLAGS.multi_gpu
Asim Shankar's avatar
Asim Shankar committed
186
      })
187

188
  # Train the model
Asim Shankar's avatar
Asim Shankar committed
189
190
191
192
  def train_input_fn():
    # When choosing shuffle buffer sizes, larger sizes result in better
    # randomness, while smaller sizes use less memory. MNIST is a small
    # enough dataset that we can easily shuffle the full epoch.
193
194
    ds = dataset.train(FLAGS.data_dir)
    ds = ds.cache().shuffle(buffer_size=50000).batch(FLAGS.batch_size).repeat(
Asim Shankar's avatar
Asim Shankar committed
195
        FLAGS.train_epochs)
196
    return ds
Asim Shankar's avatar
Asim Shankar committed
197

Asim Shankar's avatar
Asim Shankar committed
198
199
200
201
  # Set up training hook that logs the training accuracy every 100 steps.
  tensors_to_log = {'train_accuracy': 'train_accuracy'}
  logging_hook = tf.train.LoggingTensorHook(
      tensors=tensors_to_log, every_n_iter=100)
Asim Shankar's avatar
Asim Shankar committed
202
  mnist_classifier.train(input_fn=train_input_fn, hooks=[logging_hook])
203
204

  # Evaluate the model and print results
Asim Shankar's avatar
Asim Shankar committed
205
  def eval_input_fn():
206
207
    return dataset.test(FLAGS.data_dir).batch(
        FLAGS.batch_size).make_one_shot_iterator().get_next()
Asim Shankar's avatar
Asim Shankar committed
208
209

  eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn)
210
  print()
211
  print('Evaluation results:\n\t%s' % eval_results)
212

213
214
  # Export the model
  if FLAGS.export_dir is not None:
Asim Shankar's avatar
Asim Shankar committed
215
216
    image = tf.placeholder(tf.float32, [None, 28, 28])
    input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
217
        'image': image,
Asim Shankar's avatar
Asim Shankar committed
218
219
    })
    mnist_classifier.export_savedmodel(FLAGS.export_dir, input_fn)
220

221

222
223
224
225
class MNISTArgParser(argparse.ArgumentParser):

  def __init__(self):
    super(MNISTArgParser, self).__init__()
226

227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
    self.add_argument(
        '--multi_gpu', action='store_true',
        help='If set, run across all available GPUs.')
    self.add_argument(
        '--batch_size',
        type=int,
        default=100,
        help='Number of images to process in a batch')
    self.add_argument(
        '--data_dir',
        type=str,
        default='/tmp/mnist_data',
        help='Path to directory containing the MNIST dataset')
    self.add_argument(
        '--model_dir',
        type=str,
        default='/tmp/mnist_model',
        help='The directory where the model will be stored.')
    self.add_argument(
        '--train_epochs',
        type=int,
        default=40,
        help='Number of epochs to train.')
    self.add_argument(
        '--data_format',
        type=str,
        default=None,
        choices=['channels_first', 'channels_last'],
        help='A flag to override the data format used in the model. '
        'channels_first provides a performance boost on GPU but is not always '
        'compatible with CPU. If left unspecified, the data format will be '
        'chosen automatically based on whether TensorFlow was built for CPU or '
        'GPU.')
    self.add_argument(
        '--export_dir',
        type=str,
        help='The directory where the exported SavedModel will be stored.')


if __name__ == '__main__':
  parser = MNISTArgParser()
268
  tf.logging.set_verbosity(tf.logging.INFO)
269
270
  FLAGS, unparsed = parser.parse_known_args()
  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)